How does activation recomputation affect training throughput?

It reduces memory usage but increases compute time due to recomputing activations.

Checkpointing reduces peak memory footprint, enabling larger models or batch sizes, but introduces extra forward passes during backward propagation.